A fault diagnosis method for wind turbine gearbox

Abstract

The present invention relates to the field of fault diagnosis, and in particular to a fault diagnosis method for a wind turbine gearbox, comprising: collecting vibration signals of the wind turbine gearbox, dividing the collected vibration signals according to a preset length and normalizing them into fault samples; marking fault category labels on some fault samples according to historical fault reports of the wind turbine gearbox; dividing the fault samples with fault category labels into a support set, a query set, and a verification set, and taking the fault samples without fault category labels as an unlabeled sample set; constructing a fault diagnosis model, and training the fault diagnosis model through the support set, the query set, the verification set and the unlabeled sample set; acquiring a target vibration signal of the wind turbine gearbox, and normalizing the target vibration signal of the wind turbine gearbox and inputting it into a trained fault diagnosis model to calculate and obtain a fault diagnosis result. The method of the present invention can diagnose the wind turbine gearbox fault in real time.

Description

Technical Field

[0001] The invention relates to the field of fault diagnosis, and in particular to a fault diagnosis method for a gearbox of a wind turbine generator set. Background Art

[0002] The wind turbine gearbox is a key component used to increase the speed and transmit torque to the generator. It is very easy to fail under harsh and changeable working conditions, which will cause the wind turbine to shut down for maintenance and seriously affect the production plan of the wind turbine. Therefore, it is of great significance to carry out fault diagnosis of the wind turbine gearbox to ensure the safe and stable operation of the wind turbine.

[0003] Existing gearbox fault diagnosis methods based on deep learning are heavily dependent on sufficient labeled fault samples. However, in wind turbine gearbox fault diagnosis, expensive labeled fault samples are difficult to obtain, and the wind turbine gearbox fault diagnosis model trained with limited labeled fault samples is prone to overfitting, making it difficult to accurately diagnose wind turbine gearbox faults. Summary of the invention

[0004] In order to solve the problem in the prior art that it is difficult to obtain expensive labeled fault samples in wind turbine gearbox fault diagnosis, and the wind turbine gearbox fault diagnosis model trained with limited labeled fault samples is prone to overfitting, making it difficult to accurately diagnose the fault of the wind turbine gearbox, the present invention proposes a wind turbine gearbox fault diagnosis method comprising:

[0005] S1: collecting vibration signals of the wind turbine gearbox, dividing the collected vibration signals into fault samples according to preset lengths and normalizing them;

[0006] S2: Label some fault samples with fault category based on historical fault reports of wind turbine gearboxes;

[0007] S3: The fault samples with fault category labels are divided into a support set, a query set, and a validation set, and the fault samples without fault category labels are taken as an unlabeled sample set;

[0008] S4: constructing a fault diagnosis model, and training the fault diagnosis model through a support set, a query set, a validation set, and an unlabeled sample set, wherein the fault diagnosis model includes: a prototype network and a memory gated recurrent network;

[0009] Furthermore, the training of the fault diagnosis model using the support set, the query set, the validation set and the unlabeled sample set includes:

[0010] S41: pre-training the prototype network through the support set and the fault samples to be classified, and obtaining the posterior probability set of the unlabeled fault samples and the posterior probability set of the query fault samples according to the posterior probability of the fault samples to be classified in each training cycle of the prototype network, wherein the fault samples to be classified include: the fault samples in the unlabeled sample set and the fault samples in the query set;

[0011] S42: construct a pseudo-label training data set and a pseudo-label test data set according to the posterior probability set of the unlabeled fault samples and the posterior probability set of the query fault samples; and use the pseudo-label training data set to train the memory gated recurrent network, and after the training is completed, perform sample expansion on the pseudo-label test data set through the memory gated recurrent network to obtain pseudo-label fault samples and pseudo-labels of the pseudo-label fault samples, and generate a pseudo-label fault sample set;

[0012] S43: Calculate the weight of the pseudo-label fault sample using a pre-trained prototype network according to the pseudo-label fault sample and the pseudo-label of the pseudo-label fault sample;

[0013] S44: Calculate the weighted centroid class prototype of each fault category according to the pseudo-label of the pseudo-label fault sample, the fault category label, the weight of the pseudo-label fault sample, the fault samples in the support set, and the fault samples in the query set, and replace the class prototype of each fault category in the pre-trained prototype network with the weighted centroid class prototype of each fault category to obtain a semi-supervised weighted centroid prototype network;

[0014] S45: Perform momentum update on the weighted centroid prototype of the semi-supervised weighted centroid prototype network;

[0015] S46: using the pseudo-labeled fault samples and the fault samples in the query set as training samples and using the pseudo-labels of the pseudo-labeled fault samples and the fault category labels of the query fault samples to construct a loss function of the semi-supervised weighted centroid prototype network, and updating the parameters of the semi-supervised weighted centroid prototype network through back propagation to complete the training of the semi-supervised weighted centroid prototype network;

[0016] S47: Input the verification set into the trained semi-supervised weighted centroid prototype network for accuracy verification; and repeat steps S42-S47 to save the semi-supervised weighted centroid prototype network with the highest accuracy to complete the training of the fault diagnosis model.

[0017] S5: Obtain a target vibration signal of the wind turbine gearbox, normalize the target vibration signal, and input it into the trained fault diagnosis model to obtain a fault diagnosis result.

[0018] The present invention has at least the following beneficial effects:

[0019] The present invention adopts a memory-gated recurrent network to screen high-confidence unlabeled fault samples to produce pseudo-label samples, thereby avoiding the generation of erroneous pseudo-label samples; the pseudo-label samples are used to train the prototype network to enhance the stability and accuracy of the prototype network, while solving the cost of obtaining labeled fault samples, and improving the accuracy of wind turbine gearbox fault diagnosis through sufficient labeled fault samples. The semi-supervised weighted centroid prototype network proposed in the present invention is used to alleviate the prototype drift problem and improve the classification performance and stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a flow chart of the method of the present invention;

[0021] Figure 2 This is the training process of the prototype network;

[0022] Figure 3 Schematic diagram for training a memory gated recurrent network and generating a pseudo-labeled fault sample set;

[0023] Figure 4 Schematic diagram of weighted centroid prototype update strategy;

[0024] Figure 5 Schematic diagram for comparing the diagnostic accuracy of different fault diagnosis and identification methods. DETAILED DESCRIPTION

[0025] The technical solution of the present invention will be described in detail below in conjunction with the accompanying drawings of the embodiments of the present invention. The described embodiments are only part of the embodiments of the present invention and cannot be used to limit the scope of the present invention. Figure 1 As shown, a fault diagnosis method for a wind turbine gearbox includes:

[0026] S1: Collect the vibration signal of the wind turbine gearbox, divide the collected vibration signal into fault samples according to the preset length and normalize it.

[0027] Preferably, the method for collecting the vibration signal of the gearbox of the wind turbine generator set includes: collecting the vibration signal of the gearbox of the wind turbine generator set through a vibration condition monitoring system (Condition Monitoring System, CMS). For example, every day during the daily operation of the wind turbine generator set, the CMS samples the vibration signal of the gearbox of the wind turbine generator set at a sampling frequency of 25600 Hz and a sampling duration of 5.12 seconds to obtain 131072 sampled data; in order to obtain more samples, for each type of fault type, the vibration data of the gearbox five days before the fault occurs is collected, totaling 655360 data points, which are segmented and normalized with a sample length of 4096 to obtain 160 fault samples of this type of fault.

[0028] S2: Label some fault samples with fault category based on historical fault reports of wind turbine gearboxes;

[0029] Preferably, the step of labeling some fault samples with fault category labels according to the historical fault report of the wind turbine gearbox includes: labeling some fault samples with fault category labels according to the inspection report of the wind turbine gearbox and the guidance of experts, wherein the fault category labels include: the category of the fault corresponding to the fault sample, for example, gearbox bearing outer ring fault, inner ring fault, rolling element fault, etc.

[0030] S3: Divide the fault samples with fault category labels into support sets Query Set Validation set The fault samples without fault category labels are taken as the unlabeled sample set in, represents the i-th fault sample in the support set, represents the fault category label of the i-th fault sample in the support set, N S represents the number of fault samples in the support set; represents the i-th fault sample in the query set, represents the fault category label of the i-th fault sample in the query set, N Q Indicates the number of fault samples in the query set; represents the i-th fault sample in the validation set, represents the fault category label of the i-th fault sample in the validation set, N V Represents the number of fault samples in the validation set; represents the i-th fault sample in the unlabeled sample set, represents the fault category label of the i-th fault sample in the unlabeled sample set, N U Represents the number of faulty samples in the unlabeled sample set.

[0031] S4: constructing a fault diagnosis model, and training the fault diagnosis model through a support set, a query set, a validation set, and an unlabeled sample set, wherein the fault diagnosis model includes: a prototype network and a memory gated recurrent network;

[0032] The present invention constructs a fault diagnosis model using a memory gated recurrent network; the memory gated recurrent network is a method for predicting multivariate sequence data, which can not only analyze the sequence correlation of a single variable, but also analyze the intrinsic correlation between multiple variables. In view of its ability to analyze multivariate sequence data, the present invention introduces it as a sample selection model.

[0033] Preferably, the training of the fault diagnosis model using the support set, the query set, the validation set and the unlabeled sample set includes:

[0034] S41: pre-training the prototype network through the support set and the fault samples to be classified, and obtaining the posterior probability set of the unlabeled fault samples and the posterior probability set of the query fault samples according to the posterior probability of the fault samples to be classified in each training cycle of the prototype network, wherein the fault samples to be classified include: the fault samples in the unlabeled sample set and the fault samples in the query set;

[0035] See also Figure 2 Preferably, the pre-training of the prototype network by the support set and the fault samples to be classified, and obtaining the posterior probability set of the unlabeled fault samples and the posterior probability set of the query fault samples according to the posterior probability of the fault samples to be classified by the prototype network in each training cycle include:

[0036] S411: extracting feature vectors of fault samples and to-be-classified fault samples in the support set through the convolutional neural network in the prototype network; summing and averaging the feature vectors of fault samples of the same fault category in the support set according to the fault category labels, and calculating the class prototype of each fault category; the class prototype of each fault category includes:

[0037]

[0038] in, represents the i-th fault sample in the support set, CP (c) represents the class prototype of the cth fault category, N S represents the number of fault samples in the support set, express The eigenvector of , I(*) represents the indicator function.

[0039] S412: Calculate the Euclidean distance between the feature vector of the fault sample to be classified and the class prototype of each fault category, wherein the Euclidean distance between the feature vector of the fault sample to be classified and the class prototype of each fault category includes:

[0040] D (c) (x)=||f(x)-CP (c) || 2

[0041] x∈{Q∪U}

[0042] Among them, x represents the fault sample to be classified, D (c) (x) represents the class prototype CP of the fault sample x to be classified and the cth fault category (c) The Euclidean distance between them, f(x) represents the feature vector of the fault sample x to be classified, Q represents the query set, and U represents the unlabeled sample set.

[0043] S413: Calculate the posterior probability of the fault sample to be classified according to the Euclidean distance between the feature vector of the fault sample to be classified and the class prototype of each fault category, wherein the posterior probability of the fault sample to be classified includes:

[0044]

[0045] in, represents the posterior probability of the fault sample to be classified predicted by the prototype network in the e-th training cycle, p (c) (x) represents the probability that the fault sample to be classified belongs to the cth fault category in the prototype network’s eth training cycle, N represents the number of fault categories, and D (c) (x) represents the class prototype CP of the fault sample x to be classified and the cth fault category (c) The distance between them, x represents the fault sample to be classified.

[0046] S414: saving the posterior probability of the fault sample to be classified in each training cycle of the prototype network to obtain a posterior probability set of unlabeled fault samples and a posterior probability set of query fault samples;

[0047] Preferably, the posterior probability set of the unlabeled fault samples includes:

[0048]

[0049] in, represents the i-th fault sample in the query set, represents the prediction of the prototype network in the e-th training cycle The posterior probability of Indicated in the prototype network E 1 Within training cycles The posterior probability set.

[0050] The posterior probability set of the query fault sample includes:

[0051]

[0052] in, represents the i-th fault sample in the unlabeled sample set, represents the prediction of the prototype network in the e-th training cycle The posterior probability of Indicated in the prototype network E 1 Within training cycles The posterior probability set.

[0053] S415: Construct a loss function of the prototype network according to the posterior probability of querying the fault sample and the fault category label of the query fault sample in each training cycle of the prototype network, update the parameters of the prototype network through back propagation, and complete the pre-training of the prototype network, wherein the query fault sample is the fault sample in the query set, and the unlabeled sample set and the verification set are similarly named to make the present invention easier to understand.

[0054] Preferably, the loss function of the prototype network includes:

[0055]

[0056] Among them, L CE (*,*) represents the cross entropy loss function, represents the i-th fault sample in the query set, represents the fault category label of the i-th fault sample in the query set, N Q represents the number of fault samples in the query set, represents the prediction of the prototype network in the e-th training cycle The posterior probability, l PN Represents the loss function of the prototype network.

[0057] S42: constructing a pseudo-label training data set and a pseudo-label test data set according to the posterior probability set of the unlabeled fault samples and the posterior probability set of the query fault samples; and using the pseudo-label training data set to train the memory gated recurrent network, and during the training process, performing sample expansion on the pseudo-label test data set through the memory gated recurrent network to obtain pseudo-label fault samples and pseudo-labels of the pseudo-label fault samples;

[0058] Preferably, constructing a pseudo-label training data set and a pseudo-label test data set according to the posterior probability set of the unlabeled fault samples and the posterior probability set of the query fault samples comprises:

[0059] S421: sorting the posterior probability samples in the posterior probability set of the query fault sample according to the order of the prototype network training cycle; and sampling the posterior probability of the query fault sample from the posterior probability set of the query fault sample according to the preset number of samples l each time and the interval distance s between two adjacent samples to construct a pseudo-label training data set, wherein the pseudo-label training data set includes:

[0060]

[0061]

[0062] Among them, DMT represents the pseudo-label training dataset, represents the i-th fault sample in the query set, represents the fault category label of the i-th fault sample in the query set, N Q represents the number of fault samples in the query set, Indicated in the prototype network E 1 Within training cycles The posterior probability set of Indicates from The posterior probability of the bth query fault sample sampled in, Indicates from The bth query fault sample sampled in Fault category label, N H Indicates from The number of posterior probabilities of fault samples sampled for query, argmax(*) represents the maximum index function, l represents the number of samples for each sampling, s represents the interval distance between two adjacent samples, onehot(*) represents the one-hot code function, for example, the posterior probability set of query fault samples is M={A1,A2,A3,A4,A5,A6,A7,A8,A9,A10,A11,A12,A13,A14,A15, in the present invention, l takes the value of 2, indicating that two samples are sampled from M each time, s is set to 2, indicating that 2 samples are spaced between every two adjacent sampling times, then the pseudo-label training data set is {A1,A2,A5,A6,A9,A10,A13,A14, as Figure 3 As shown, A1 represents the posterior probability predicted by the prototype network for the query fault sample in the first training cycle, A2, A3... and so on.

[0063] S422: Construct a pseudo-label test data set according to the posterior probability set of the unlabeled fault samples, wherein the pseudo-label test data set includes:

[0064]

[0065] Among them, DMP represents the pseudo-label test data set, represents the fault samples in the unlabeled sample set, N U represents the number of fault samples in the unlabeled sample set, Unlabeled samples Faulty samples The posterior probability set.

[0066] Preferably, the training of the memory gated recurrent network using the pseudo-label training data set includes:

[0067] The pseudo-label training data set is used as a training sample and the loss function of the memory gated recurrent network is constructed according to the fault category label of the query fault sample. The parameters of the memory gated recurrent network are updated by back propagation to complete the training of the memory gated recurrent network. The loss function of the memory gated recurrent network includes:

[0068]

[0069] Positive sample

[0070] Negative samples

[0071] in, represents the prediction of the prototype network in the e-th training cycle The posterior probability of , argmax(*) represents the maximum index function, and onehot(*) represents the one-hot encoding function; when and At the same time, is defined as a negative sample, otherwise it is defined as a positive sample, N H Indicates from The number of posterior probabilities of sampled query fault samples in , N Q represents the number of fault samples in the query set, L MSE (*,*) represents the mean square error loss function, Indicates from The posterior probability of the bth query fault sample sampled in, Indicates from The bth query fault sample sampled in Fault category label, N H+ Represents the number of positive samples in the pseudo-label training dataset, N H- represents the number of negative samples in the pseudo-label training dataset, Represents the memory gated recurrent network pair The logical output of , λ represents the positive and negative sample balance factor parameter, represents the positive and negative sample balance factor, l PPE represents the loss function of the memory gated recurrent network, and N represents the number of fault categories.

[0072] Preferably, the method of performing sample expansion on the pseudo-label test data set through the trained memory gated recurrent network to obtain pseudo-label fault samples and pseudo-labels of pseudo-label fault samples includes:

[0073] The posterior probability of the unlabeled fault samples in the pseudo-labeled test dataset is input into the trained memory-gated recurrent network to predict the category and decision factor of the posterior probability of the unlabeled fault samples, and the pseudo-labeled fault samples and the pseudo-labels of the pseudo-labeled fault samples are generated according to the category and decision factor of the posterior probability of the pseudo-unlabeled fault samples.

[0074] Preferably, the categories and decision factors of the posterior probability of the unlabeled fault samples include:

[0075]

[0076] in, represents the fault sample in the unlabeled fault sample set, represents the posterior probability of the unlabeled fault samples in the pseudo-labeled test dataset, Represents the memory gated recurrent network pair The forward propagation of Indicates that the memory gated recurrent network is trained in the eth cycle The category prediction of represents the e-th training cycle of the memory gated recurrent network The decision factor of is, and N represents the number of fault categories.

[0077] Preferably, generating a pseudo-labeled fault sample and a pseudo-label of the pseudo-labeled fault sample according to the category and decision factor of the posterior probability of the pseudo-unlabeled fault sample comprises:

[0078] The posterior probability of the unlabeled fault samples whose decision factors are less than the set threshold is used as the pseudo-label fault samples, and the category prediction of the pseudo-label fault samples by the memory gated recurrent network is used as the pseudo-label of the pseudo-label fault samples to generate a pseudo-label fault sample set. Preferably, the pseudo-label fault sample set includes:

[0079]

[0080] Among them, PLDS e represents the pseudo-label fault sample set, represents the i-th pseudo-label fault sample obtained in the e-th training cycle of the memory gated recurrent network, express The pseudo-label of represents the i-th unlabeled fault sample in the unlabeled sample set U, Indicates that the memory gated recurrent network is trained in the eth cycle The category prediction of represents the e-th training cycle of the memory gated recurrent network The decision factor prediction of represents the number of pseudo-label fault samples obtained in the e-th training cycle of the memory gated recurrent network, argmax(*) represents the maximum index function, represents the posterior probability of the unlabeled faulty samples in the pseudo-labeled test dataset.

[0081] S43: Calculate the weight of the pseudo-label fault sample using a pre-trained prototype network according to the pseudo-label fault sample and the pseudo-label of the pseudo-label fault sample;

[0082] Preferably, the calculating the weight of the pseudo-label fault sample using a pre-trained prototype network according to the pseudo-label fault sample and the pseudo-label of the pseudo-label fault sample comprises:

[0083] S431: Input the pseudo-label fault sample into the pre-trained prototype network to perform forward propagation on the pseudo-label fault sample and calculate the pseudo-loss of the pseudo-label fault sample according to the pseudo-label of the pseudo-label fault sample. The pseudo-loss of the pseudo-label fault sample includes:

[0084]

[0085] Among them, PLDS represents the pseudo-label fault sample set, represents the i-th pseudo-label fault sample obtained in the e-th training cycle of the memory gated recurrent network; Represents the pre-trained prototype network pair The category prediction of express The pseudo label, L CE (*,*) represents the cross entropy loss function, Represents a pseudo-label fault sample The pseudo loss, represents the set of pseudo losses of all pseudo-label fault samples in the pseudo-label fault sample set PLDS, represents the number of pseudo-labeled fault samples obtained in the e-th training cycle of the memory gated recurrent network.

[0086] S432: Calculate the weight of the pseudo-label fault sample according to the pseudo-loss of the pseudo-label fault sample, where the weight of the pseudo-label fault sample includes:

[0087]

[0088] in, represents the maximum value of the pseudo losses of all pseudo-labeled fault samples, represents the minimum value of the pseudo loss of all pseudo-labeled fault samples, Represents a pseudo-label fault sample The pseudo loss, Represents a pseudo-label fault sample The weight of represents the i-th pseudo-label fault sample obtained in the e-th training cycle of the memory gated recurrent network.

[0089] See also Figure 4, S44: Calculate the weighted centroid class prototype of each fault category according to the pseudo-label of the pseudo-label fault sample, the fault category label, the weight of the pseudo-label fault sample, the fault samples in the support set and the fault samples in the query set, and replace the class prototype of each fault category in the pre-trained prototype network with the weighted centroid class prototype of each fault category to obtain a semi-supervised weighted centroid prototype network. Since pseudo-label fault samples with feature distribution differences from the support set fault samples are introduced in the prototype calculation process, the class prototype will fluctuate to a certain extent in different training cycles, which is defined as prototype drift; the prototype drift problem will directly affect the classification performance and stability of the prototype network; for this reason, the present invention proposes a semi-supervised weighted centroid prototype network to alleviate the prototype drift problem;

[0090] Preferably, the weighted centroid class prototype of each fault category includes:

[0091]

[0092]

[0093] Among them, CP WC represents the weighted centroid class prototype of the semi-supervised weighted centroid prototype network, The weighted centroid class prototype of the c-th fault class, represents the i-th fault sample in the support set, represents the fault category label of the i-th fault sample in the support set, N S represents the number of fault samples in the support set, represents the i-th pseudo-label fault sample obtained in the e-th training cycle of the memory gated recurrent network; express The pseudo-label of represents the number of pseudo-label fault samples, Represents a pseudo-label fault sample The weight of express The eigenvector of Represents the convolutional neural network extracted by the prototype network The eigenvector of , I(*) represents the indicator function.

[0094] S45: Perform momentum update on the weighted centroid prototype of the semi-supervised weighted centroid prototype network;

[0095] The method of updating the weighted centroid prototype of the semi-supervised weighted centroid prototype network includes:

[0096]

[0097] Where η is the momentum factor, which is set to 0.99. represents the weighted centroid class prototype of the e-th training cycle of the semi-supervised weighted centroid prototype network, Represents the weighted centroid class prototype of the e-1th training cycle of the semi-supervised weighted centroid prototype network.

[0098] S46: Taking the pseudo-label fault samples and the fault samples in the query set as training samples and using the pseudo-label fault samples' pseudo-labels and the fault category labels of the query fault samples to construct a loss function of the semi-supervised weighted centroid prototype network, and updating the parameters of the semi-supervised weighted centroid prototype network through back propagation to complete the training of the semi-supervised weighted centroid prototype network, wherein the loss function of the semi-supervised weighted centroid prototype network includes:

[0099]

[0100] Among them, l SSWCPN represents the loss function of the semi-supervised weighted centroid prototype network, β represents the adjustment factor parameter which can be randomly selected by a person skilled in the art, and in the present invention, the value is 0.5, represents the i-th fault sample in the query set, express Fault category label, N S represents the number of fault samples in the support set, represents the i-th pseudo-label fault sample obtained in the e-th training cycle of the memory gated recurrent network; express The pseudo-label of represents the number of pseudo-label fault samples, It represents the semi-supervised weighted centroid prototype network in the e-th training cycle The posterior probability of It represents the semi-supervised weighted centroid prototype network in the e-th training cycle The posterior probability of .

[0101] S47: Input the verification set into the trained semi-supervised weighted centroid prototype network for accuracy verification; and repeat steps S42-S46 to save the semi-supervised weighted centroid prototype network with the highest accuracy to complete the training of the fault diagnosis model.

[0102] Preferably, inputting the verification set into the trained semi-supervised weighted centroid prototype network for accuracy verification comprises:

[0103]

[0104] Among them, N V represents the number of faulty samples in the validation set, I(*) represents the indicator function, and Accuracy represents the accuracy of the semi-supervised weighted centroid prototype network.

[0105] S5: Obtain a target vibration signal of the wind turbine gearbox, normalize the target vibration signal, and input it into the trained fault diagnosis model to obtain a fault diagnosis result.

[0106] Preferably, the target vibration signal of the wind turbine gearbox is a wind turbine gearbox vibration signal collected in real time by the CMS;

[0107] The target vibration signal of the wind turbine gearbox is normalized and then input into a trained fault diagnosis model to obtain a fault diagnosis result, which includes: inputting the target vibration signal of the wind turbine gearbox into a saved semi-supervised weighted centroid prototype network with the highest accuracy for fault diagnosis, and outputting a diagnosis result.

[0108] In order to verify the effectiveness and advancement of the wind turbine gearbox fault diagnosis method disclosed in the present invention, a comparative test was conducted on the vibration signal data of the wind turbine gearbox; as shown in the following table, the vibration signal data has a total of 6 different wind turbine gearbox health states, each of which contains 1216 samples (each sample contains 4096 data points).

[0109] Wind turbine gearbox vibration signal data description

[0110]

[0111] Four groups of experiments with different small sample sizes were set up. In each group of experiments, samples of a certain size were randomly selected as labeled training samples, 500 samples were selected as test samples, and the rest were discarded with label information as unlabeled samples.

[0112] The specific allocation of each data set is shown in the following table:

[0113] Distribution of fault sample data sets for each type of health status

[0114]

[0115] Based on the above sample sets, the semi-supervised weighted centroid prototype network SSWCPN proposed in this paper is compared with seven other methods, including CNN, SN described in the paper ICML.,215. "Siamese neural networks for one-shot image recognition.", PN described in the paper NIPS.,217. "Prototypical networks for few-shot learning.", PLL described in the paper arXiv preprint:206.05278,Jun,22. "An overview of deep semi-supervised learning.", LST described in the paper ICLR.,Feb,218. "Meta-learning for semi-supervised few-shot classification.", TPN described in the paper ICIP.,Oct,22. "Transductive prototypical network for few-shot classification.", and RRPN described in the paper ICISC.,Nov.22. "Robust re-weighting prototypical networks for few-shot classification."

[0116] To improve the fairness and persuasiveness of the experiment, the compared methods were subjected to five independent experiments based on the same experimental parameters, data, and feature extractors. The specific experimental results are as follows: Figure 5 And the following table shows:

[0117] Experimental Results

[0118]

[0119] It is not difficult to see that among all the methods, CNN has the worst fault diagnosis effect, and only achieved 46.07%, 46.07%, 55.95% and 62.47% recognition accuracy in each group of experiments, which is far from meeting the needs of fault diagnosis. Due to the advantages in dealing with small sample problems, the fault diagnosis results of SN and PN have been improved to a certain extent. PN has improved by 25.12%, 22.24%, 22.07% and 2.13% respectively compared with CNN in each group of experiments; and because PLL uses CNN for fault diagnosis, in addition, since PLL cannot obtain satisfactory classification results in the PT stage, PLL's classification effect on unlabeled samples has also become worse, which makes it difficult for PLL to learn from pseudo-labels with noise. Therefore, PLL has only made some progress over SN, but is still better than CNN, improving by 14.43%, 16.73%, 18.18% and 2.80% respectively in various experiments. Compared with PLL, the FSSSL method uses PN as the network backbone, which is more suitable for small problems than CNN, and uses proportional sample selection to improve the accuracy of the pseudo-labeled samples used, so it has better results than PLL. Among them, LST only introduces pseudo-label samples to optimize the parameters of the feature embedding module instead of using them to refine the category prototype, so the performance of LST is not as good as TPN and RRPN; compared with TPN, when using pseudo-label samples to adjust the category prototype, RRPN notices the adverse effect of incorrect pseudo-label samples on the generation of category prototypes, and further adopts a re-weighting mechanism to solve this problem. Therefore, the robustness of the category prototype is enhanced, and RRPN obtains the best fault diagnosis performance among existing methods, with an accuracy of 74.94%, 83.21%, 89.79% and 94.75% in various experiments.

[0120] In addition, this paper further extends the research of FSSSL and develops SSWCPN. SSWCPN first designs a sample selection strategy based on MGRN to enhance the reliability and accuracy of pseudo-labeled samples. This strategy does not require manual setting of proportions or thresholds, and can select confident samples through historical posterior probabilities; then, SSWCPN proposes a WCCP prototype update strategy based on weighted mechanism and momentum algorithm to solve the prototype drift problem. These two improvements enable SSWCPN to obtain more stable and accurate category prototypes, and enable SSWCPN to obtain the best fault diagnosis results in this comparative experiment. In various experiments, the fault diagnosis accuracy of SSWCPN reached 79.78%, 91.67%, 94.53% and 96.72%, respectively, and outperformed the existing best method RRPN by 4.84%, 8.46%, 4.74% and 1.97%.

[0121] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described with reference to the preferred embodiments of the present invention, it should be understood by those skilled in the art that various changes may be made in form and details without departing from the spirit and scope of the present invention as defined in the appended claims.

Claims

1. A fault diagnosis method for a wind turbine gearbox, characterized in that: include: S1: collecting vibration signals of the wind turbine gearbox, dividing the collected vibration signals into fault samples according to preset lengths and normalizing them; S2: Label some fault samples with fault category based on historical fault reports of wind turbine gearboxes; S3: The fault samples with fault category labels are divided into a support set, a query set, and a validation set, and the fault samples without fault category labels are taken as an unlabeled sample set; S4: constructing a fault diagnosis model, and training the fault diagnosis model through a support set, a query set, a validation set, and an unlabeled sample set, wherein the fault diagnosis model includes: a prototype network and a memory gated recurrent network; The training of the fault diagnosis model by using the support set, the query set, the validation set and the unlabeled sample set includes: S41: pre-training the prototype network through the support set and the fault samples to be classified, and obtaining the posterior probability set of the unlabeled fault samples and the posterior probability set of the query fault samples according to the posterior probability of the fault samples to be classified in each training cycle of the prototype network, wherein the fault samples to be classified include: the fault samples in the unlabeled sample set and the fault samples in the query set; S42: construct a pseudo-label training data set and a pseudo-label test data set according to the posterior probability set of the unlabeled fault samples and the posterior probability set of the query fault samples; and use the pseudo-label training data set to train the memory gated recurrent network, and after the training is completed, perform sample expansion on the pseudo-label test data set through the memory gated recurrent network to obtain pseudo-label fault samples and pseudo-labels of the pseudo-label fault samples, and generate a pseudo-label fault sample set; S43: Calculate the weight of the pseudo-label fault sample using a pre-trained prototype network according to the pseudo-label fault sample and the pseudo-label of the pseudo-label fault sample; S44: Calculate the weighted centroid class prototype of each fault category according to the pseudo-label of the pseudo-label fault sample, the fault category label, the weight of the pseudo-label fault sample, the fault samples in the support set, and the fault samples in the query set, and replace the class prototype of each fault category in the pre-trained prototype network with the weighted centroid class prototype of each fault category to obtain a semi-supervised weighted centroid prototype network; S45: Perform momentum update on the weighted centroid prototype of the semi-supervised weighted centroid prototype network; S46: using the pseudo-labeled fault samples and the fault samples in the query set as training samples and using the pseudo-labels of the pseudo-labeled fault samples and the fault category labels of the query fault samples to construct a loss function of the semi-supervised weighted centroid prototype network, and updating the parameters of the semi-supervised weighted centroid prototype network through back propagation to complete the training of the semi-supervised weighted centroid prototype network; S47: input the verification set into the trained semi-supervised weighted centroid prototype network for accuracy verification; and repeat steps S42-S47 to save the semi-supervised weighted centroid prototype network with the highest accuracy, thereby completing the training of the fault diagnosis model; S5: Obtain a target vibration signal of the wind turbine gearbox, normalize the target vibration signal, and input it into the trained fault diagnosis model to obtain a fault diagnosis result.

2. A fault diagnosis method for a wind turbine gearbox according to claim 1, characterized in that: The prototype network is pre-trained by using the support set and the fault samples to be classified, and the posterior probability set of the unlabeled fault samples and the posterior probability set of the query fault samples are obtained according to the posterior probability of the fault samples to be classified in each training cycle of the prototype network, including: S411: extracting feature vectors of fault samples in the support set and fault samples to be classified through the convolutional neural network in the prototype network; summing and averaging the feature vectors of fault samples of the same fault category in the support set according to the fault category labels, and calculating the class prototype of each fault category; S412: Calculate the Euclidean distance between the feature vector of the fault sample to be classified and the class prototype of each fault category; S413: Calculating the posterior probability of the fault sample to be classified according to the Euclidean distance between the feature vector of the fault sample to be classified and the class prototype of each fault category; S414: saving the posterior probability of the fault sample to be classified in each training cycle of the prototype network to obtain a posterior probability set of unlabeled fault samples and a posterior probability set of query fault samples; S415: constructing a loss function of the prototype network according to the posterior probability of querying the fault sample and the fault category label of the query fault sample in each training cycle of the prototype network, updating the parameters of the prototype network through back propagation, and completing the pre-training of the prototype network.

3. A fault diagnosis method for a wind turbine gearbox according to claim 1, characterized in that: The step of constructing a pseudo-label training data set and a pseudo-label test data set according to the posterior probability set of the unlabeled fault samples and the posterior probability set of the query fault samples includes: S421: sorting the posterior probability samples in the posterior probability set of the query fault samples according to the order of the prototype network training cycle; and constructing a pseudo-label training data set by sampling the posterior probabilities of the query fault samples from the posterior probability set of the query fault samples according to the preset sampling number l and sampling interval s; S422: Construct a pseudo-label test data set based on the posterior probability set of unlabeled fault samples.

4. A fault diagnosis method for a wind turbine gearbox according to claim 1, characterized in that: The pseudo-label fault samples and pseudo-label fault samples obtained by performing sample expansion on the pseudo-label test data set through the trained memory gated recurrent network include: The posterior probability of the unlabeled fault samples in the pseudo-label test dataset is input into the trained memory gated recurrent network to predict the category and decision factor of the posterior probability of the unlabeled fault samples; The posterior probability of the unlabeled fault samples whose decision factors are less than the set threshold is taken as the pseudo-labeled fault samples, and the category prediction of the pseudo-labeled fault samples by the memory gated recurrent network is taken as the pseudo-label of the pseudo-labeled fault samples.

5. A wind turbine gearbox fault diagnosis method according to claim 1 or 4, characterized in that: The pseudo-label fault sample set includes: Among them, PLDS e represents the pseudo-label fault sample set, represents the i-th pseudo-label fault sample obtained in the e-th training cycle of the memory gated recurrent network, express The pseudo-label of represents the i-th unlabeled fault sample in the unlabeled sample set U, Indicates that the memory gated recurrent network is trained in the eth cycle The category prediction of represents the e-th training cycle of the memory gated recurrent network The decision factor prediction of represents the number of pseudo-label fault samples obtained in the e-th training cycle of the memory gated recurrent network, argmax(*) represents the maximum index function, represents the posterior probability of the unlabeled faulty samples in the pseudo-labeled test dataset.

6. A fault diagnosis method for a wind turbine gearbox according to claim 1, characterized in that: The method of calculating the weight of the pseudo-label fault sample using a pre-trained prototype network according to the pseudo-label fault sample and the pseudo-label of the pseudo-label fault sample includes: S431: Input the pseudo-label fault sample into the pre-trained prototype network to perform forward propagation on the pseudo-label fault sample and calculate the pseudo loss of the pseudo-label fault sample according to the pseudo-label of the pseudo-label fault sample; S432: Calculate the weight of the pseudo-label fault sample according to the maximum and minimum values ​​of the pseudo losses of all pseudo-label fault samples.

7. A wind turbine gearbox fault diagnosis method according to claim 1, characterized in that: The weighted centroid class prototypes for each fault category include: Among them, CP WC represents the weighted centroid class prototype of the semi-supervised weighted centroid prototype network, The weighted centroid class prototype of the c-th fault class, represents the i-th fault sample in the support set, represents the fault category label of the i-th fault sample in the support set, N S represents the number of fault samples in the support set, represents the i-th pseudo-label fault sample obtained in the e-th training cycle of the memory gated recurrent network; express The pseudo-label of represents the number of pseudo-label fault samples, Represents a pseudo-label fault sample The weight of express The eigenvector of Represents the convolutional neural network extracted by the prototype network The eigenvector of , I(*) represents the indicator function.

8. A fault diagnosis method for a wind turbine gearbox according to claim 1, characterized in that: The loss function of the semi-supervised weighted centroid prototype network includes: Among them, l SSWCPN represents the loss function of the semi-supervised weighted centroid prototype network, β represents the adjustment factor parameter, represents the i-th fault sample in the query set, express Fault category label, N S represents the number of fault samples in the support set, represents the i-th pseudo-label fault sample obtained in the e-th training cycle of the memory gated recurrent network; express The pseudo-label of represents the number of pseudo-label fault samples, It represents the semi-supervised weighted centroid prototype network in the e-th training cycle The posterior probability of It represents the semi-supervised weighted centroid prototype network in the e-th training cycle The posterior probability of .

Images